Adaptiv modellering och realtidsinlärning med många okända variabler.
Tidsperiod: 2013-01-01 till 2016-12-31
Projektledare: Kristiaan Pelckmans
Budget: 3 287 000 SEK
This project will advance theoretical insights in techniques of handling large sets of unknowns in methods of adaptive modeling and online learning. Within this context, focus will be given to the class of Gradient Descent (GD) flavored methods. Recent literature witnessed a renewal of interest in those due to their robustness, computational convenience and potential effectiveness using proper tuning. Dealing with large sets of unknowns is a common theme in related field as compressive sensing, but is only covered sporadically in a context of adaptive modeling or online learning. However, technical tools for such are around, and it is the aim of this project to piece those together into theoretically effective algorithms. Such body of results will culminate into a reference work on the general topic, and into a range of application studies in (i) adaptive compression, (ii) wireless sensor networks and (iii) automatic control and sequential design. Such theoretical and conceptual results will be verified and validated in a context of medical analysis and bio-informatics. Especially, we intend to validate results in studies of Mass Spectrometry Imaging (MSI) and Genome Wide Association (GWA). Those two cases are found to have an urgent need for such tools as modern experimentation devices produce collections of data which are hardly manageable using existing tools.